jiev/few-shot-adversarial-learning-for-face-swap

This is a unofficial re-implementation of the paper "Few-Shot Adversarial Learning of Realistic Neural Talking Head Models"

Few-Shot-Adversarial-Learning-for-face-swap

This is a unofficial re-implementation of the paper "Few-Shot Adversarial Learning of Realistic Neural Talking Head Models" based on Pytorch.

description

The paper from SAMSUNG AI lab presents a new novel and efficient method for face swap, which has amazing performance. I am a student and interested in this paper, so I try to reproduce it. I have writen the script for getting landmarks dataset, and dataloader pipeline. I also construct the whole network including Embedding, Generator and Discriminator with loss functions. Everything is done according to the paper. But, due to unspecific network descripted by the paper, I found that there are some mistakes in Generator, especially for the Adaptive instance normalization. The training process gets weird results, which you can look at it from "training_visual" file. For getting more understanding about this amazing and great work, I open source this projection and invite more people who are interested in it to become contributor. If someone reproduces it successfully in the future, please tell me. Thanks!

how to use

1. get landmarks information.

you should download the VoxCeleb, about 36GB. And after unzip the dataset, please change the name of root path of the dataset as "voxcelb1", or you can adapt "./face-alignment_projection/get_landmarks.py" for your envirnoment.

python get_landmarks.py

For test you shouldn't preprocess all data in the VoxCeleb dataset, because it's time-consuming. I just use about 200 video clips. The two files are generated. "video_clips_path.txt" records all clips which have been preprocessed. "video_landmarks_path.txt" records landmarks information path.

2. download VGGFace weights for perceptual loss.

you can download from here on this website. make sure that put the weight file under the "./pretrained" file path.

3. start to train the whole network.

python ./train.py

For each 100 iteartion, you can look at temp results including fake image, GT and landmarks under the "./training_visual" file path.

Cite

[1] Few-Shot Adversarial Learning of Realistic Neural Talking Head Models

[2] face-alignment is from the great work